crop done

stefanklut · Dec 12, 2023 · dc2e462 · dc2e462
1 parent d0dfa26
commit dc2e462
Show file tree

Hide file tree

Showing 2 changed files with 168 additions and 48 deletions.
diff --git a/configs/segmentation/baseline/baseline_ovdr_crop.yaml b/configs/segmentation/baseline/baseline_ovdr_crop.yaml
@@ -0,0 +1,108 @@
+# python main.py --config configs/segmentation/baseline_ovdr.yaml --tmp_dir /encrypted/tmp
+_BASE_: ../segmentation_base.yaml
+
+OUTPUT_DIR: /encrypted/baseline
+
+NAME: ovdr_scaling_test
+
+PREPROCESS:
+  OVERWRITE: False
+  DISABLE_CHECK: False
+
+  RESIZE:
+    RESIZE_MODE: "scaling"
+    SCALING: 0.5
+    MAX_SIZE: -1
+
+  BASELINE:
+    LINE_WIDTH: 10
+
+INPUT:
+  RESIZE_MODE: "scaling"
+  SCALING_TRAIN: 1.0
+  MAX_SIZE_TRAIN: -1
+
+  MAX_SIZE_TEST: -1
+
+  FORMAT: RGB
+
+  GRAYSCALE:
+    PROBABILITY: 0.1
+
+  BRIGHTNESS:
+    PROBABILITY: 0.2
+    MIN_INTENSITY: 0.5
+    MAX_INTENSITY: 1.5
+
+  CONTRAST:
+    PROBABILITY: 0.2
+    MIN_INTENSITY: 0.5
+    MAX_INTENSITY: 1.5
+
+  SATURATION:
+    PROBABILITY: 0.1
+    MIN_INTENSITY: 0.5
+    MAX_INTENSITY: 1.5
+
+  GAUSSIAN_FILTER:
+    PROBABILITY: 0.2
+    MIN_SIGMA: 0.5
+    MAX_SIGMA: 1.5
+
+  HORIZONTAL_FLIP:
+    PROBABILITY: 0.
+
+  VERTICAL_FLIP:
+    PROBABILITY: 0.
+
+  ELASTIC_DEFORMATION:
+    PROBABILITY: 0.2
+    ALPHA: 0.1
+    SIGMA: 0.01
+
+  AFFINE:
+    PROBABILITY: 1.
+
+    TRANSLATION:
+      PROBABILITY: 0.9
+      STANDARD_DEVIATION: 0.02
+
+    ROTATION:
+      PROBABILITY: 0.9
+      KAPPA: 30.
+
+    SHEAR:
+      PROBABILITY: 0.9
+      KAPPA: 20.
+
+    SCALE:
+      PROBABILITY: 0.9
+      STANDARD_DEVIATION: 0.12
+
+  CROP:
+    ENABLED: True
+    SIZE: [1024, 1024]
+    TYPE: absolute
+
+SOLVER:
+  IMS_PER_BATCH: 2
+  CHECKPOINT_PERIOD: 25000
+  BASE_LR: 0.0002
+  GAMMA: 0.1
+  STEPS: () #(80000, 120000, 160000)
+  MAX_ITER: 500000
+
+MODEL:
+  MODE: baseline
+
+  SEM_SEG_HEAD:
+    NUM_CLASSES: 2
+
+  WEIGHTS:
+
+TRAIN:
+  WEIGHTS:
+
+TEST:
+  WEIGHTS:
+  EVAL_PERIOD: 10000
diff --git a/datasets/augmentations.py b/datasets/augmentations.py
@@ -951,35 +951,36 @@ def build_augmentation(cfg: CfgNode, mode: str = "train") -> list[T.Augmentation
     if not mode == "train":
         return augmentation
 
-    # Color augments
-    augmentation.append(RandomApply(Grayscale(image_format=cfg.INPUT.FORMAT), prob=cfg.INPUT.GRAYSCALE.PROBABILITY))
-    augmentation.append(
-        RandomApply(
-            RandomBrightness(
-                intensity_min=cfg.INPUT.BRIGHTNESS.MIN_INTENSITY,
-                intensity_max=cfg.INPUT.BRIGHTNESS.MAX_INTENSITY,
-            ),
-            prob=cfg.INPUT.BRIGHTNESS.PROBABILITY,
-        )
-    )
+    # Moving pixels
+
     augmentation.append(
         RandomApply(
-            RandomContrast(
-                intensity_min=cfg.INPUT.CONTRAST.MIN_INTENSITY,
-                intensity_max=cfg.INPUT.CONTRAST.MAX_INTENSITY,
+            RandomAffine(
+                t_stdv=cfg.INPUT.AFFINE.TRANSLATION.STANDARD_DEVIATION,
+                r_kappa=cfg.INPUT.AFFINE.ROTATION.KAPPA,
+                sh_kappa=cfg.INPUT.AFFINE.SHEAR.KAPPA,
+                sc_stdv=cfg.INPUT.AFFINE.SCALE.STANDARD_DEVIATION,
+                probabilities=(
+                    cfg.INPUT.AFFINE.TRANSLATION.PROBABILITY,
+                    cfg.INPUT.AFFINE.ROTATION.PROBABILITY,
+                    cfg.INPUT.AFFINE.SHEAR.PROBABILITY,
+                    cfg.INPUT.AFFINE.SCALE.PROBABILITY,
+                ),
             ),
-            prob=cfg.INPUT.CONTRAST.PROBABILITY,
+            prob=cfg.INPUT.AFFINE.PROBABILITY,
         )
     )
+
     augmentation.append(
         RandomApply(
-            RandomSaturation(
-                intensity_min=cfg.INPUT.SATURATION.MIN_INTENSITY,
-                intensity_max=cfg.INPUT.SATURATION.MAX_INTENSITY,
+            RandomElastic(
+                alpha=cfg.INPUT.ELASTIC_DEFORMATION.ALPHA,
+                sigma=cfg.INPUT.ELASTIC_DEFORMATION.SIGMA,
             ),
-            prob=cfg.INPUT.SATURATION.PROBABILITY,
+            prob=cfg.INPUT.ELASTIC_DEFORMATION.PROBABILITY,
         )
     )
+
     augmentation.append(
         RandomApply(
             RandomGaussianFilter(
@@ -1010,53 +1011,64 @@ def build_augmentation(cfg: CfgNode, mode: str = "train") -> list[T.Augmentation
         )
     )
 
+    # Orientation
+
     augmentation.append(
         RandomApply(
-            RandomElastic(
-                alpha=cfg.INPUT.ELASTIC_DEFORMATION.ALPHA,
-                sigma=cfg.INPUT.ELASTIC_DEFORMATION.SIGMA,
+            RandomOrientation(
+                orientation_percentages=cfg.INPUT.ORIENTATION.PERCENTAGES,
             ),
-            prob=cfg.INPUT.ELASTIC_DEFORMATION.PROBABILITY,
+            prob=cfg.INPUT.ORIENTATION.PROBABILITY,
         )
     )
 
+    # Color augments
     augmentation.append(
         RandomApply(
-            RandomAffine(
-                t_stdv=cfg.INPUT.AFFINE.TRANSLATION.STANDARD_DEVIATION,
-                r_kappa=cfg.INPUT.AFFINE.ROTATION.KAPPA,
-                sh_kappa=cfg.INPUT.AFFINE.SHEAR.KAPPA,
-                sc_stdv=cfg.INPUT.AFFINE.SCALE.STANDARD_DEVIATION,
-                probabilities=(
-                    cfg.INPUT.AFFINE.TRANSLATION.PROBABILITY,
-                    cfg.INPUT.AFFINE.ROTATION.PROBABILITY,
-                    cfg.INPUT.AFFINE.SHEAR.PROBABILITY,
-                    cfg.INPUT.AFFINE.SCALE.PROBABILITY,
-                ),
+            Grayscale(
+                image_format=cfg.INPUT.FORMAT,
             ),
-            prob=cfg.INPUT.AFFINE.PROBABILITY,
+            prob=cfg.INPUT.GRAYSCALE.PROBABILITY,
         )
     )
-
     augmentation.append(
         RandomApply(
-            RandomOrientation(
-                orientation_percentages=cfg.INPUT.ORIENTATION.PERCENTAGES,
+            RandomBrightness(
+                intensity_min=cfg.INPUT.BRIGHTNESS.MIN_INTENSITY,
+                intensity_max=cfg.INPUT.BRIGHTNESS.MAX_INTENSITY,
             ),
-            prob=cfg.INPUT.ORIENTATION.PROBABILITY,
+            prob=cfg.INPUT.BRIGHTNESS.PROBABILITY,
+        )
+    )
+    augmentation.append(
+        RandomApply(
+            RandomContrast(
+                intensity_min=cfg.INPUT.CONTRAST.MIN_INTENSITY,
+                intensity_max=cfg.INPUT.CONTRAST.MAX_INTENSITY,
+            ),
+            prob=cfg.INPUT.CONTRAST.PROBABILITY,
+        )
+    )
+    augmentation.append(
+        RandomApply(
+            RandomSaturation(
+                intensity_min=cfg.INPUT.SATURATION.MIN_INTENSITY,
+                intensity_max=cfg.INPUT.SATURATION.MAX_INTENSITY,
+            ),
+            prob=cfg.INPUT.SATURATION.PROBABILITY,
         )
     )
 
-    # augmentation.append(RandomApply(RandomTranslation(t_stdv=cfg.INPUT.AFFINE.TRANSLATION.STANDARD_DEVIATION),
-    #                                 prob=cfg.INPUT.AFFINE.PROBABILITY * cfg.INPUT.AFFINE.TRANSLATION.PROBABILITY))
-    # augmentation.append(RandomApply(RandomRotation(r_kappa=cfg.INPUT.AFFINE.ROTATION.KAPPA),
-    #                                 prob=cfg.INPUT.AFFINE.PROBABILITY * cfg.INPUT.AFFINE.ROTATION.PROBABILITY))
-    # augmentation.append(RandomApply(RandomShear(sh_kappa=cfg.INPUT.AFFINE.SHEAR.KAPPA),
-    #                                 prob=cfg.INPUT.AFFINE.PROBABILITY * cfg.INPUT.AFFINE.SHEAR.PROBABILITY))
-    # augmentation.append(RandomApply(RandomScale(sc_stdv=cfg.INPUT.AFFINE.SCALE.STANDARD_DEVIATION),
-    #                                 prob=cfg.INPUT.AFFINE.PROBABILITY * cfg.INPUT.AFFINE.SCALE.PROBABILITY))
+    # Crop
+    if cfg.INPUT.CROP.ENABLED:
+        augmentation.append(
+            RandomCrop_CategoryAreaConstraint(
+                crop_type=cfg.INPUT.CROP.TYPE,
+                crop_size=cfg.INPUT.CROP.SIZE,
+                single_category_max_area=cfg.INPUT.CROP.SINGLE_CATEGORY_MAX_AREA,
+            )
+        )
 
-    # print(augmentation)
     return augmentation